Re: Public SPARQL endpoints:managing (mis)-use and communicating limits to users.

On 4/18/13 7:06 PM, Andrea Splendiani wrote:
> Il giorno 18/apr/2013, alle ore 16:04, Kingsley Idehen <kidehen@openlinksw.com> ha scritto:
>
>> On 4/18/13 9:23 AM, Andrea Splendiani wrote:
>>> Hi,
>>>
>>> I think that some caching with a minimum of query rewriting would get read of 90% of the select{?s ?p ?o} where {?s?p ?o} queries.
>> Sorta.
>> Client queries are inherently unpredictable. That's always been the case, and that predates SPARQL. These issues also exist in the SQL RDBMS realm, which is why you don't have SQL endpoints delivering what SPARQL endpoints provide.
> I know, but I suspect that these days lot of these "intensive" queries are explorative, just to check what is in the dataset, and may end up being very similar in structure.

Note, we have logs and recordings of queries that hit many of our public 
endpoints. For instance, we are preparing a report on DBpedia that will 
actually shed light on types and complexity of queries that hit the 
DBpedia endpoint.

> Jerven: can you report on your experience in this ? How much of problematic queries are not really targeted, but more generic ?
>
>>>  From a user perspective, I would rather have a clear result code upfront telling me: your query is to heavy, not enough resources and so on, than partial results + extra codes.
>> Yes, and you get that in some solutions e.g., what we provide. Basically, our server (subject to capacity) will tell you immediately that your query exceeds the query cost limits (this is different from timeout limits). The aforementioned feature was critical to getting the DBpedia SPARQL endpoint going, years ago.
> Can you make a precise estimation of the query cost, or do you rely on some heuristics ?

We have a query cost optimizer. It handles native and distributed 
queries. Of course, query optimization is a universe onto itself, but 
over the years we've continually applied what we've learned about 
queries into its continued evolution.

>
>>> I won't do much of partial results anyway... so it's time wasted both sides.
>> Not in a world where you have a public endpoint and zero control over the queries issued by clients.
>> Not in a world where you to provide faceted navigation over entity relations as part of a "precision find" style service atop RDF based Linked Data etc..
> I mean, partial results are ok if I have control on which part it is... a system-dependent random subset of results is not much useful (not even for statistics!)

You have control with our system because you are basically given the 
ability to retry using a heuristic that increases the total processing 
time per retry, at the same time, while you are making up your mind 
about whether retry or not, there are background activities running in 
relation your last query. Remember, query processing is comprised of 
many parts:

1. parsing
2. costing
3. solution
4. actual retrieval.

Many see 1-4 as a monolith. Not so when dealing with DBMS processing. 
Again, this is novel, its quite old in the RDBMS realm.

>
>>> One empiric solution could be to assign a quota per requesting IP (or other form of identification).
>> That's but one coarse-grained factor. You need to be able to associate a user agent (human or machine) profile with what ever quality of service you seek to scope to said profile. Again, this is the kind of thing we offer by leveraging WebID, Inference, and RDF right inside the core DBMS engine.
> I agree. The finer the better. The IP-based approach is perhaps relatively easy to implement if not much is provided by the system.
>
>>>   Then one could restrict the total amount of resource per time-frame, possibly with smart policies.
>> "Smart Policies" are the kind of thing you produce by exploiting the kind or entity relationship semantics baked into RDF based Linked Data. Basically, OWL (which is all about describing entity types and relation types semantics) serves this purpose very well. We certainly put it to use in our data access policy system which enables us to offer different capabilities and resource consumption to different human- or machine-agent profiles.
> How do you use OWL for this ?

We just have normal RDF graphs that describe data access policies. All 
you need is a Linked Data URI that denotes an Agent (human or machine) , 
agent profile oriented ontology (e.g., FOAF and the like) that defines 
entity types and relation types, a Web access control ontology, actual 
entity relations based on the aforementioned ontologies, and reasoning 
capability.  Basically, using RDF to do what it's actually designed to do.

>
>>> It would also avoid people breaking big queries in many small ones...
>> You can't avoid bad or challenging queries. What you can do is look to fine-grained data access policies (semantically enhanced ACLs) to address this problem. This has always been the challenge, even before the emergence of the whole Semantic Web , RDF etc.. The same challenges also dogged the RDBMS realm. There is no dancing around this matter when dealing with traditional RDBMS or Web oriented data access.
>>> But I was wondering: why is resource consumption a problem for sparql endpoint providers, and not for other "providers" on the web ? (say, YouTube, Google, ...).
>>> Is it the unpredictability of the resources needed ?
>> Good question!
>>
>> They hide the problem behind airport sized data centers, and then they get you to foot the bill via your profile data which ultimately compromises your privacy.
> Isn't the same possible with sparql, in principle ?

Sorta.

The ultimate destination, in our opinion is this, which setup provides 
the most cost-effective solution for Linked Data exploitation, at 
Web-scale. Basically, how many machines do you need to provide 
acceptable performance to a variety of user and agent profiles. We don't 
believe you need an airport sized data center to pull that off. The LOD 
cloud cache is just an 12-node Virtuoso cluster split across 4 machines.

"OpenLink Virtuoso version 07.00.3202, on Linux 
(x86_64-unknown-linux-gnu), Cluster Edition(12 server processes, 756 GB 
total memory)"

That's at the footer of the system home page: http://lod.openlinksw.com. 
Likewise, we expose timing and resource utilization data per query 
processed via that interface.

> Although, I guess if a company would know that you spy on their queries... there would be some issue (unlike for users and facebook, for some reason).

It works like this:

1. we put out a public endpoint
2. we allow the public certain kinds of access e.g., like the DBpedia 
fair use policy we have in place
3. we can provide special access to specific agents based on data access 
policy graphs scoped to their WebIDs or other types of identifiers.

Kingsley


>
> best,
> Andrea
>
>> This is a problem, and it's the ultimately basis for showcasing what RDF (entity relationship based data model endowed with *explicit* rather than *implicit* human- and machine-readable entity relationship semantics) is actually all about.
>>
>>
>> Kingsley
>>> best,
>>> Andrea
>>>
>>> Il giorno 18/apr/2013, alle ore 12:53, Jerven Bolleman <jerven.bolleman@isb-sib.ch> ha scritto:
>>>
>>>> Hi All,
>>>>
>>>> Managing a public SPARQL endpoint has some difficulties in comparison to managing a simpler REST api.
>>>> Instead of counting api calls or external bandwidth use we need to look at internal IO and CPU usage as well.
>>>>
>>>> Many of the current public SPARQL endpoints limit all their users to queries of limited CPU time.
>>>> But this is not enough to really manage (mis) use of an endpoint. Also the SPARQL api being http based
>>>> suffers from the problem that we first send the status code and may only find out later that we can't
>>>> answer the query after all. Leading to a 200 not OK problem :(
>>>>
>>>> What approaches can we come up with as a community to embedded resource limit exceeded exceptions in the
>>>> SPARQL protocols. e.g. we could add an exception element to the sparql xml result format.[1]
>>>>
>>>> The current limits to CPU use are not enough to really avoid misuse. Which is why I submitted a patch to
>>>> Sesame that allows limits on memory use as well. Although limits on disk seeks or other IO counts may be needed by some as well.
>>>>
>>>> But these are currently hard limits what I really want is
>>>> "playground limits" i.e. you can use the swing as much as you want if you are the only child in the park.
>>>> Once there are more children you have to share.
>>>>
>>>> And how do we communicate this to our users. i.e. this result set is incomplete because you exceeded your IO
>>>> quota please break up your queries in smaller blocks.
>>>>
>>>> For my day job where I do manage a 7.4 billion triple store with public access some extra tools in managing users would be
>>>> great.
>>>>
>>>> Last but not least how can we avoid that users need to run SELECT (COUNT(DISTINT(?s) as ?sc} WHERE {?s ?p ?o} and friends.
>>>> For beta.sparql.uniprot.org I have been moving much of this information into the sparql endpoint description but its not a place
>>>> where people look for this information.
>>>>
>>>> Regards,
>>>> Jerven
>>>>
>>>> [1] Yeah these ideas are not great timing just after 1.1 but we can always start SPARQL 1.2 ;)
>>>>
>>>>
>>>>
>>>> -------------------------------------------------------------------
>>>> Jerven Bolleman                        Jerven.Bolleman@isb-sib.ch
>>>> SIB Swiss Institute of Bioinformatics      Tel: +41 (0)22 379 58 85
>>>> CMU, rue Michel Servet 1               Fax: +41 (0)22 379 58 58
>>>> 1211 Geneve 4,
>>>> Switzerland     www.isb-sib.ch - www.uniprot.org
>>>> Follow us at https://twitter.com/#!/uniprot
>>>> -------------------------------------------------------------------
>>>>
>>>>
>>>
>>>
>>>
>>
>> -- 
>>
>> Regards,
>>
>> Kingsley Idehen	
>> Founder & CEO
>> OpenLink Software
>> Company Web: http://www.openlinksw.com
>> Personal Weblog: http://www.openlinksw.com/blog/~kidehen
>> Twitter/Identi.ca handle: @kidehen
>> Google+ Profile: https://plus.google.com/112399767740508618350/about
>> LinkedIn Profile: http://www.linkedin.com/in/kidehen
>>
>>
>>
>>
>>
>
>
>
>


-- 

Regards,

Kingsley Idehen	
Founder & CEO
OpenLink Software
Company Web: http://www.openlinksw.com
Personal Weblog: http://www.openlinksw.com/blog/~kidehen
Twitter/Identi.ca handle: @kidehen
Google+ Profile: https://plus.google.com/112399767740508618350/about
LinkedIn Profile: http://www.linkedin.com/in/kidehen